How does cuda code work when rasterization is performed ?

[zcc00210]

Hi,
Thanks a lot for sharing your great work. I have some questions about the process of differentiable gaussian rasterization.
1.

What's the meaning of "from . import _C" ? Is "_C" refer to the gaussian rasterization folder containing cuda code?
2.

Which file in the above rasterize_gaussians folder containing cuda code should I look for the corresponding rasterize_gaussians function when I perform rasterization, running the code to the line num_rendered, color, objects, radii, geomBuffer, binningBuffer, imgBuffer = _C.rasterize_gaussians(*args)?
3.

There are too many code files in this folder, do they have any order of execution, I don't know where to start reading them.
And how is this cuda code connected to the corresponding parts of python, is it by the way in question 1?
4. I am new in this field. I would like to add a new attribute to the Gaussian and modify the code in the rasterization section accordingly, which file in cuda would you recommend to start with?

Sorry for asking so many questions, looking forward to your reply！
My best regards

[yuanyibo666]

Same question

[cmh1027]

As you can see in ext.cpp, rasterize_gaussians function in python is connected to RasterizeGaussiansCUDA in rasterize_points.cu.
The order of execution is rasterize_points.cu => raterizer_impl.cu => forward.cu (backward.cu)

[AsherJingkongChen]

My recommendation is to look through all the files in the repository, especially *.cu. You need some math supplements (linear transformation, matrix, eigenvalues, partial differentiation) when reading the uncommented parts.

The backward part is useless if you are not interested in training.